[CINN] Add castOp before addOp by AddCastToElementwiseAddPass

[DrRyanHuang]

PR Category

Execute Infrastructure

PR Types

Bug fixes

Description

运行 GPT OSS 20B 出现：

----------------------
Error Message Summary:
----------------------
InvalidArgumentError: The operands' types of the node [Add] don't match. 
Received types: float32 and bfloat16
  [Hint: Expected a.type() == b.type(), but received a.type():float32 != b.type():bfloat16.] 
(at /workspace/Paddle/paddle/cinn/ir/ir.cc:138)

打出IR图：

    (%544) = "cinn_op.fusion" [id:12442] () -> tensor<-1x32xf32> {
		...
        (%548) = "pd_op.add" [id:12441] (%543, %547) {stop_gradient:[false]} : (tensor<-1x32xf32>, tensor<-1x-1xf16>) -> tensor<-1x32xf32> ......
        () = "cf.yield" [id:12443] (%548) {} : (tensor<-1x32xf32>) ->  {  }
    } { ...... }

是这里的 pd_op.add 出现了问题，本PR添加 AddCastToElementwiseAddPass 在 add op 之前插入 cast op，保证类型一致

--

这个 PR 给 add 的两个输入都添加了 cast ，如下：

Paddle/paddle/cinn/hlir/dialect/operator/transforms/add_cast_to_elementwise_add_pass.cc

Lines 78 to 93 in f956a3b

	if (x_dtype != y_dtype) {
	pir::Type output_dtype = GetOutputDtype(x_dtype, y_dtype);
	auto cast_op0 = rewriter.Build<paddle::dialect::CastOp>(
	op->operand_source(0),
	::paddle::dialect::TransToPhiDataType(output_dtype));
	auto cast_op1 = rewriter.Build<paddle::dialect::CastOp>(
	op->operand_source(1),
	::paddle::dialect::TransToPhiDataType(output_dtype));
	op->operand(0).set_source(cast_op0->result(0));
	op->operand(1).set_source(cast_op1->result(0));
	return true;
	}

不过不必担心 CreateAddBroadcastToElementwisePass 是 ApplyGroupOpPass 中的第一个Pass
之后其中的不必要的 cast 将会被其他 pass 消除

Paddle/paddle/cinn/hlir/dialect/operator/transforms/add_cinn_pass.cc

Lines 195 to 201 in f956a3b

	void ApplyGroupOpPass(::pir::Program* program,
	const std::function<std::shared_ptr<pir::PassManager>()>&
	CreatePassManager) {
	std::shared_ptr<pir::PassManager> pass_manager = CreatePassManager();
	pass_manager->AddPass(
	cinn::dialect::ir::CreateAddBroadcastToElementwisePass());

这个 Pass 处于 CINN 前端：

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 89.58333% with 5 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@99a2725). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...tor/transforms/add_cast_to_elementwise_add_pass.cc	89.36%	5 Missing ⚠️

❌ Your patch status has failed because the patch coverage (89.58%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #76359   +/-   ##
==========================================
  Coverage           ?   89.58%           
==========================================
  Files              ?        2           
  Lines              ?       48           
  Branches           ?        0           
==========================================
  Hits               ?       43           
  Misses             ?        5           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[DrRyanHuang]

TODO: 后续本PR内容需要与此处的💩代码一起删除
PS: 请查看 history ，💩这个 emoji 不是我写的，是 @SigureMo 加的
PSS: 啥 history 呀，我不到啊

Paddle/paddle/phi/common/type_promotion.h

Lines 206 to 222 in c645437

	if (x_dtype != y_dtype) {
	// TODO(Xi Zhao): we got special case for add now, should remove it in future.
	#ifdef PADDLE_WITH_CUDA
	if ((op_name == "add" || op_name == "add_") &&
	x_dtype == DataType::FLOAT32 &&
	(y_dtype == phi::DataType::BFLOAT16 ||
	y_dtype == phi::DataType::FLOAT16)) {
	return false;
	}
	#elif defined(PADDLE_WITH_XPU)
	if ((op_name == "add" || op_name == "add_") &&
	x_dtype == DataType::FLOAT32 &&
	(y_dtype == phi::DataType::BFLOAT16 ||
	y_dtype == phi::DataType::FLOAT16)) {
	return false;
	}
	#endif